CAC-v0.1
CAC is a large language model with 6.7B parameters specifically finetuned for code completions. CAC, although trained for code autocompletion, can also be used for other code related tasks such as:
- Generation
- Summarization
- Translation
- Question Answering
- Optimization
- Debugging
- Code Review and more.
This is the very first version of CAC (0.1) and is still under development. For this version, we chose to go ahead with DeepSeek-Coder-6.7B as the base model.
Model Details
-
Training Data: Exclusively fine-tuned on a proprietary dataset of 1.8 billion tokens of high-quality programming problems and solutions.
-
The dataset was generated manually and is internal to CodeMate.
-
Training Techniques: The model was fine-tuned using Flash Attention 2.
-
A sequence length of 8096 tokens was used during training.
-
Multilingual Support: CAC-v0.1 is proficient in multiple programming languages, including Python, C/C++, TypeScript, Java, and more.
Load the model with Transformers:
Make sure to install Transformers from the main git branch:
pip install git+https://github.com/huggingface/transformers.git
How to Prompt the Model:
This model accepts prompts in the Alpaca/Vicuna instruction format. For example:
### System Prompt
You are an intelligent programming assistant.
### User Message
Implement a linked list in C++
### Assistant
...
You can also use the Mistral chat template for conversations:
<s>[INST] .... [/INST] ... </s>
Load the model:
from transformers import AutoTokenizer, AutoModelForCausalLM
# Initialize the model
model_path = "codemateai/CodeMate-v0.1"
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_path)
# ... generate response ...
Limitations
This model has undergone very limited testing. CodeMate recommends additional safety testing before any real-world deployments.
For more information and updates, visit the CodeMate website.